The Discussion: Generative AI - from Consumer Rights to Human Rights #19

Welcome to The Future of You. This week I'm in conversation with two people who closely assess, monitor, and certainly in one case deploy AI in their everyday work. I talk with Josh Monk, who is the commercial director of retail and consumer at Faculty, one of the UK's best known AI and decision intelligence companies, and Sam Gregory, executive director at Witness, the human rights organization. I thought this would make a great combination, or perhaps I should say comparison, because what I discussed with Josh is the effect of AI on consumers, and with Sam, the effect of AI on citizens. Although, of course, there is crossover now, because as I would argue, in a digital world, the boundaries between consumer and citizen are fast disappearing. Look at it like this: once everything is a service, then everyone is a subscriber, no longer just a citizen or just a consumer. But maybe that's a thought for another day. For now, let me just say that this episode covers a lot of ground. Generative models, what they are and how they're changing the customer experience, testing and iterating as a process if you're working with them, and how the field of prompt engineering is going to evolve. Also, shallowfakes versus deepfakes, the novel ability to create new media content from existing data, the need for safeguards and transparency, and of course, what's next for GPT. And the next generation of models. But first, is this where we all expected to be in 2023? Josh, welcome to the show. Now, the last time we spoke was a while ago, back in 2020, when you very kindly helped me with a lot of research for my book, and we were talking about generative AI and its implications. Way back then, and we will return to prompt engineering in a moment and what you had to say about it then. But I wondered if we could start by you telling me if this is where you expected us to be in 2023. So we've had GPT exploding onto the scenes, disrupting everything from the creative industries to education systems. We've had Italy banning its use, of course, as well as some legal firms lately. We've had the "Pause" letter, which we may come onto, and we've had the AI-generated song by Drake and The Weeknd, that was not by Drake and the Weeknd actually at all. So presumably we're only at the beginning of all of this, or are we? I'd love to get your take.

Great question. And I think, you know, even for those of us that have been kind of in the field for a while and thinking about AI and generative AI, I think many of us have been caught off guard by the pace that this technology has exploded into the public awareness. And it's hard to kind of sometimes remember that it was only, you know, less than 6 months ago that ChatGPT was kind of released into the public domain and how quickly that gained so many users. So I, you know, I thought we would have made some good progress since 2020, and GPT-3, when it was originally released, was a big step up in terms of capability and performance. But the pace that we've seen over the last 6 months, the amount of stuff that has happened, the amount of progress that's been made, but also the amount of awareness of this set of technologies and its capabilities has moved far more quickly in the last 6 months than I thought it might do. And it's, it's interesting to think about how much of that is down to the technology itself, how much of that is down to interfaces to the technology and, and making these sorts of models more accessible for everyday people to interact with and, and, and get a feel for? And how much is due to, you know, regulators and legislators starting to wake up to the, you know, possible impact that these sorts of things are going to have on everyday life and the world around us?

Yeah, that's so interesting to hear people in the industry saying that even they are in part surprised about the speed of this. And I guess it's the scale as well, isn't it? I mean, the scale of GPT and what OpenAI have done, not to mention the cost, of course, but the trillions, I don't know, it must be more than that, of words that have gone into this training data. I mean, just the sheer scale. Can it continue to get bigger and bigger, or have we seen almost the limits now of the large language model? Are we into sort of, you know, reinforcement learning or new ways of presenting this?

Yeah, I mean, that is a good question because I think up until now we've certainly seen a kind of trend where bigger is better, and GPT-4 is certainly much bigger and certainly much better than, than GPT-3.5 or ChatGPT. But actually it may be that we're starting to kind of hit a plateau of the improvements that you can get from just adding more data, more words into these models, and actually that we see changes to the architecture of the models or the reinforcement learning and human feedback steps that sit on top of those models are actually the next wave of improvement. So that's how we're going to get the next, the next level of improvements of these systems. And certainly I know, like, Sam Altman has pretty much said as much fairly recently that we may kind of be topping out on just adding more data to GPT models. Others in the— other influencers in the community have got maybe more fundamental challenges with GPT models altogether as a route to generalized intelligence. But yeah, I think the answer is probably a bit of both. Like, I think probably there is still value to be gleaned from from bigger, but it might not be quite as linear as we've seen historically. And actually, we see maybe more of the improvements in the future coming from other approaches being added on top of those large models that really generate the next level of performance.

I was going to ask you, actually, Faculty have got a partnership of some sort with OpenAI, right? You should tell us about that. But also, what do you think about the market as a whole? Obviously, we've had Google coming back with Bard and other competitors. And as you say, they might be slightly behind, or maybe they're not. I don't know. It'd be really nice to have an overview from an insider about, about the market overall.

I think the market is in a really interesting space at the moment. As you said, Faculty are partnered with OpenAI, and our objective from the work that we do with our customers is all around helping them to use AI in a productive way and in a safe way. And so for us, when we were looking at the market, the right choice was to partner with the OpenAI guys who we saw as having both the best capability but also the kind of most robust and safe capabilities when it comes to actually using these in a commercial or enterprise context. But, you know, the reality is it is quite a volatile market. There are other players that are doing some really amazing stuff. The team at Anthropic, you mentioned the team at Google within the image space, you've got Stability and Midjourney. And recently Hugging Face released, released their open source kind of chat, large language chat-based large language model. So it's moving really quickly. I think overall that's good. It's a good thing. We're seeing that having some, you know, spurring on innovation and through competition and also pricing. And we've seen some kind of dramatic shifts and drops in things like pricing, which if you're a business looking to actually use this stuff to do significant amounts of work in your business, that makes a difference and it opens up use cases that otherwise might be, might be prohibitive. As to how it shakes out, you know, I can take a guess at the next 6 to 12 months, but beyond that, beyond that, it's kind of anyone's guess. It does seem like, you know, to some extent there's two directions that this could go. One, which is kind of consolidation, where the requirements for data size, data volumes, training costs essentially make it prohibitive for anyone but a very small number of vendors to have models that are truly like viable, you usable large language models or generalized models in a sense that people could use. I think it's also, you know, interesting to think that not every use case, not every business needs the biggest large language model for everything. And so I think it's quite feasible that we see a, you know, a bit of a tiered market where you have a small number of operators at the top of the pyramid offering these very large models, but then actually a much wider number of players offering small large language models, but for more, maybe more specialized or use-case-specific applications.

Yes, that's fascinating. And you hinted at that when we spoke about this in 2020, actually, that there would be like services starting up, you know, to help companies with very bespoke ways to integrate this, which kind of brings us on nicely to the question I wanted to ask you, which was, you know, how are businesses operationalizing these transformers, if indeed they are? Because that's one of the questions, is anybody doing it well? Is everybody experimenting? How are businesses and organizations doing that? I know you work with a lot, a lot of them, so perhaps you could give us a little bit of insight into that. Sure.

Yeah.

So everyone's experimenting is the thing I'd say first. I think there's very few executive teams that have not at least had the conversation about how might this be relevant to them, either in a sense of a threat to how they run currently or as an opportunity of how they could operate. So every business I speak to is starting to think about this. Some are further forward and have started to deploy these things for particular use cases. And I think what we're seeing very much is this idea of let's pick one or two places where we see a really strong fit between what the models can do and what we need the use case to be able to do. And so the types of applications we're seeing there are places where One, it's a, it's a process that involves language, right? Because these are language models. And so in, in general, they're going to be processes that involve large amounts of, um, language. So things like customer service or, or marketing where language is a key part of the process. But also too, really importantly, where there is an element of human in the loop or validation around that. And, and yes, I think we are seeing kind of a a push towards, you know, things like chatbots and fully automated services. But in general, we are seeing most businesses taking a slightly more thoughtful approach than just fully automating their customer experience using a large language model and actually saying we can use these to augment the way that our current teams operate, to make them more efficient, to make them be able to respond more quickly, or to be able to provide more context and make them better performing in the way that they respond or interact with customers. The power of these models to digest large amounts of data and then support a human customer service agent, for example, is actually tremendous and is a business case all on its own. You don't need to fully automate your entire customer service team, and it probably is not the right time to make that level of relatively risky investment before you validated the technology yourself. So yeah, I think lots of use cases that we're seeing are in that kind of, you know, customer engagement, customer understanding, customer service kind of space. And generally in a way in which is kind of augmentative rather than fully automated.

Now, I know you, you work a lot in the area of retail, retail and consumer, don't you? So can you tell us some of the types of consumer interactions or customer experiences that might be transformed by, by some of this tech?

Yeah, I think this is one of the, a really profound moment actually, this kind of the birth of these sorts of generative models is really profound in retail and consumer. And I think this is actually not a set of industries or sectors that is new to AI and machine learning and data science. It's, you know, relatively common now for retailers to use AI machine learning methods to support with things like product search or marketing personalization. But I think what's changed is the dynamic from kind of these more traditional narrow AI models to these generative models. So we used to think of models or machine learning as kind of moving up this pyramid of descriptive to predictive to prescriptive. So better understanding what happened in the past to predicting what is going to happen in the future, to then telling me what the best action is to take and maybe automating that. And ultimately, for a kind of a retail or consumer sense, a lot of what that allowed you to do was, was filter, filter and sort and rank things. So you've got a lot of products that you could show to a customer. What's the best one I should show them? You have a number of offers or of marketing messages that you could send to someone, what's the best one to, to send to them? With generative AI, it kind of sits outside of that framework a little bit because no longer are you just filtering and ranking and sorting, you're generating new data. And so I think that really transforms how you might think about how you interact with a shop or a consumer company or your network operator or really anyone where you have a kind of customer provider relationship with, because rather than that organization offering you a number of different experiences or products and then the models filter that, actually you might have an agent or a conversational AI that is actually able to generate something entirely new and bespoke and novel for you, both in the way you search for it, interact with it, and potentially even in the product offering itself. And so I think we're seeing the start of that, but That's going to take a little while for these sorts of businesses to reconfigure themselves, if indeed some of them can, moving from traditional retail view of the world to one which is totally bespoke, maybe even at the customer level.

I've got so many questions now I want to ask you about that. Where shall I start? Well, one of them is, will this replace the whole idea of segmentation? Clients love their very, very expensive segmentation models, and some of them are useful and some of them aren't, I think. But That's what is kind of, or was, the industry standard, you know, a quite finessed segmentation model. What you're explaining there is a completely different model, and you don't go anywhere near segmentation because you don't need to because you have personalization. So, how will this evolve for companies, do you think?

Yeah, you know, segmentation is such an interesting one because some versions of the world segmentation ends up being the tail that wags the dog, and you see businesses get so religious almost about their segmentation that they kind of forget that it's surely only a means to an end of serving a customer in a slightly more nuanced or granular way. And so I think, yeah, it's a really good question. We've seen, we've seen the evolution of kind of segmentation go from broad personas— male, 25 to 35, based in London— down to more granular kind of behavioral based upon the types of activities you like to take through to kind of segment of one. But all of that is really just a means to an end, right? Which is how do I better personalize an interaction and a homepage, a message in a world in which that interface is now mediated by a model entirely. Maybe you don't need to have that segmentation. And actually, the model carries a lot of the weight of that personalization work. So rather than you, the retailer, having an incredibly granular segmentation that covers every single nuance and variety of customer behavior, and then you use that to personalize your customer interactions for each individual person. Actually, the model plays that role of understanding exactly what the customer needs, exactly what they're looking for, exactly how they like to purchase, and then pulls the right set of products, of offers, of messages from the bank or generates them entirely to generate an experience that is truly personalized and totally bypasses that idea of having a, you know, very granular, rigid segmentation on which a decision is made. So that's a fundamental shift. That's just not how retailers think today. But you can imagine it being transformative when it comes to how a customer experience looks from now to 10 years in the future.

What about the people that don't want to interface, if you like, with companies like this? —so somebody who perhaps doesn't have a smartphone, doesn't want to speak to a chatbot, etc. Going back to like the segmentation psychology, they don't have to all be old people. Like, I think there's this thought that like, "Old people don't want to use this tech." I see it all over the place, you know, some people do want to get really engaged in this, and other people want the almost purely human interaction. What happens for the people that fall outside of the potential analysis by the model?

Yeah, I think basically the good retailers will crack that nut and realize that personalization is not like hyper-personalizing every customer's experience. It's actually realizing that some customers don't want that. And that, you know, actually the right thing to do is to provide the balance, like use these things as tools that augment rather than a total replacement for how you did things in the past. That need and requirement for nuance in customer relationships is key. And I think we'll see differences play out across different sectors and different industries. You know, the luxury sector, you know, the relationship that customers have with a brand like Chanel, for example, is fundamentally different from, let's say, Tesco, right? They're both fantastic brands, but just the nature of the relationships and interactions that customers have with those two businesses is different. And then within those brands, the nature of the different customers, different customer relationships have will be fundamentally different. So I think the retailers and the brands that approach this thoughtfully will, will bake that into their plan and will avoid alienating kind of large, important portions of their customer base by forcing everyone into the same working kind of operating paradigm. But it just seems inevitable that there will be this kind of gradual change. And even for things like, you know, cashless and contactless payment, right? You know, we've seen a massive adoption and uptake of that. Over the last 10 to 15 years, but some people still want to use cash. There are retailers that have been possibly burned and found problems with the fact that they've kind of lost demand because they've not supported that. Others have made the kind of the trade-off, the business case that actually that's worth it. The costs of cash handling is not worth it.

And that's going to be case by case. So again, that's back to having the human insight and understanding of the brand and how its customers want to engage, isn't it?

Outside of the model, really. Yeah, I don't think we're yet at the point where we kind of say, you know, GPT, what's our business strategy for the next 3 years?

I'm sure some people have plugged that in.

I'm sure they have, but I don't know that I would want to run my business on that quite yet.

No, exactly. I wanted to ask you about how this technology interfaces with other technology in a retail space. So, for example, over the past few years, we've seen a lot of kind of, you know, try-on, particularly in beauty, skincare, things like that, and obviously clothing. You know, try it on, even, you know, with hairstyles and anything that's kind of cosmetic like that. How does that data feed in, if it does, to a sort of an LLM or a personalization model that we've been discussing here? Do they integrate?

I think they very quickly will start to. And actually, you know, we've got so excited about large language models over the last 4 or 5 months that we've kind of forgotten that really the, the space is more broadly about generative models, and that kind of imagery and video is, is a really big part of that. You know, it wasn't that long ago that we were talking about Midjourney and DALL-E and some of these amazing, amazing models. And, you know, the next way, the next version of GPT will have kind of multimodal capability, and we'll see a lot tighter coupling between text and images, images and audio, audio and text. And, and so that's exciting. So you know, what does that mean for how the customer experience looks like for things like trying on or visualization of clothing or furniture or designs? I think really exciting. I think really, really exciting. Over the last few weeks, we've been trying to decorate our dining room, which is the fastest path to an argument in my house, is kind of interior decoration. And we were using GPT to kind of do some brainstorming around color schemes and furniture types and then took the outputs of that and put it into DALL-E to kind of generate some sample kind of examples of what that could look like. And that's kind of very rough and ready, quite rudimentary, but you don't have to think very hard at all to imagine how that process could look in a more industrialized way when you shop online for furniture or for interior decoration, or when you search indeed for clothing, where actually the process becomes far more personalized, far more individualized at a visual level, and you're able to really design and create and co-create what you want something to look like before you ever even try it on and maybe even before it even exists. And it's kind of— it's possible to imagine a world in which you're no longer picking from a catalog of products, but actually what's created is actually made for you based upon that design process, that co-design process that is totally enabled by generative models which take the form of image and text.

Yeah, I think there's a sustainability argument that people apply to that where you're literally only producing a range of one, um, the thing that you chose to bring into reality from a, from a virtual space, right? Which is again just fascinating, really.

And also on the sustainability piece, I think it's interesting as well because there is the kind of, what if we, what if we just produce what we have demand for? But also, what if we can eliminate, you know, the things like returns and orders that don't, that don't go ahead because people have had the chance to really genuinely see how this item will look or see how the, the thing will fit in a really, really visual, clear, precise, high-fidelity way before they ever buy it. So I think clearly there's a, there's a bit of a metaverse kind of question to that as well, and like quite how that is mediated is, is, is is a separate question, but I think really interesting how the, the shopping experience changes could have kind of profound sustainability impacts as well in terms of the things that we need to produce to satisfy potential demand.

So over the course of this podcast, in previous episodes, we've talked about the death of authenticity and the new era of pro-felicity. And Hans-Georg Müller was very good on this, and I've used his work in papers I've written. To make the case that we are now as identities in a digital era in cyberspace, really the summation of our profiles. So actually the information that is collected on us and the way we are seen by the tech platforms has by default become our identity in a sense. Are these generative AI models going to add to that, do you think? Do they fuel even further profilicity and the death of authenticity, do you think?

I think augment, you know, they augment, and it's hard to, it's hard to come down on, on one side too strongly or another. I think, you know, this idea of identity is, is so interesting as it pertains to the consumer. And I'm certainly not the first person to say you are what you buy, which is something which is increasingly true of people today, but it's pretty different from generations. Even, you know, our parents' generations would maybe not have strongly, strongly defined themselves by the choice they make about the clothes they wear, the food they eat, the holidays they go on. But increasingly, that is an increasing— that is a really important part of how we view ourselves. I think what's really changing with generative models is that whereas in the past you had systems that could filter down a set number of options and kind of recommend the one that might be the right fit for us, in a generative world, you might have systems that entirely co-create an experience, an outfit, a meal just for us. And at that point, when you are who you buy, in some sense your personality is kind of deeply interconnected, interconnected with that generation process. So you're no longer just using a, a dumb machine learning system to filter out the things that you don't want and buy the things that you do. The system is playing a really important role in helping you actually decide and create what you decide to purchase, whatever that is. And I think In that paradigm, in that world, that idea of identity then kind of gets wrapped up in this agent or this model, and that presents a really interesting question as to who we are as it pertains to what we decide to consume.

So will they really know who we are, or will they just know what we tend to buy and predict what we might buy next, and therefore just some of the patterns that pertain to us rather than actually like our name and who we are as a person.

Yeah, I, I think at some point the difference between knowing who we are actually and knowing deeply enough our personal choices and preferences around everything we buy, the places we go, may become a bit indistinguishable. Because at a certain point you can represent someone well enough by the behavioral choices around how they decide to spend their money that knowing their name or their address at that point kind of becomes a superfluous piece of detail. And that's really what I'm talking about. If that model is so deeply in tune with the way you like to purchase and consume, you know, media and clothing and, and the like, at what point does it become really an extension of you versus just an agent? That's an interesting question that I don't know the answer to.

I wanted to talk to you about prompt engineering and where you think we are with that. Obviously, when I talked to you back in 2020, Tracey, you talked to me about prompt engineering. Well, I didn't really know what it was then, of course. And you were saying that we might find GPT analysts and GPT specialists who exist purely to help organizations tune and optimize their in-house versions of these models across a variety of use cases. We may see organizations or individuals that specialize in producing content, such as marketing, specifically targeted for consumption of these algorithms. At the moment, people are all talking about prompt engineering, but there's another group of people saying, "Oh, that's just a stepping stone to something else." How important do you think it is that people understand, can do, can carry out prompt engineering to a high degree?

I don't know, honestly. I really don't know. It sounds like I was kind of being quite prescient back in 2020, talking about that stuff, but actually, I'm sure I was just basically parroting some wisdom that I had read somewhere else at the time. But yeah, it is really interesting. And clearly at the moment there is, there are a load of people doing some extremely interesting stuff when it comes to how you interact with these models in effective ways. And there's really interesting examples of that. There are people that are doing that in the kind of red teaming side and finding out if you ask these models in clever ways, they'll reveal things in ways that they were not supposed to, potentially harmful information. That's a form of prompt engineering. There are people that are doing things like asking the models to think step by step. And they find that if you do that, they find that actually you tend to get better, more well-reasoned, more high-quality responses out of these models. And then of course, there's the kind of prompt engineering, which is, you know, setting these models up to give them a clearer sense of what type of output you want, what type of response looks good, and what type of response looks bad, and how they should evaluate the inputs and and how they should then generate outputs. So all of that kind of prompt engineering, I think, is great. And actually, I've been amazed with how quickly, again, my Twitter feed has filled up with people with amazing examples of kind of clever and creative prompts that have been produced. Again, I'm sure the OpenAI guys have talked a little bit about that as well and said actually that space of people engineering prompts has moved way more quickly than they thought. And actually, I think probably the best prompt engineers potentially don't even work for one of these companies, the large language model companies, or kind of developers. They're external. How does the field evolve from here? Like I said, I don't know. One of the things I would say is over time there is clearly a need to, to kind of iterate on prompts. And for the businesses that are using large language models, they may find that actually continuing to optimize the prompts over time is one of the ways that they continue to make sure that the models can work very, very well. And, you know, there's lots of things that are happening in that space to allow the ongoing optimization of prompts to a point where maybe the prompts themselves are not even human English. They're just kind of like vectors, like in some embedding space that are being continually optimized, that generate the best high-quality outputs. And at that point, you're not even prompting, you're just feeding a model a string of numbers because, you know, this is the string of numbers that tends to generate the best search results or response to a customer query. So how does that work over the course of the next few years? And to what degree is it important that the information that goes into the language model is extremely easy for a person to understand as long as the outputs are good enough? We'll have to see. That might be something that's very use case dependent where actually in some cases it's fine. In other cases, no, we really, really need to know what was asked. Asked or what was requested, such that we can be confident, we can be confident on the outputs.

And so, um, how do you think AI collaboration in this space is going to work? Um, I was going to say in the workplace, but they're not workplaces anymore, are they? Because we're not in workplaces. Well, we are, but they're not just the office. Do you have any thoughts on this about the new skills? I mean, because you've kind of touched on it a little bit, the new kinds of ways that the modes or the skills that we will need to adopt in order to integrate this technology into what we're doing.

I think one of the really interesting things that you can see right away is this idea of kind of co-collaboration, co-pilot, co-creation with these models. And so the skill that that really necessitates is, is kind of thinking out loud, isn't it? It's an ability to work in a way where you make an initial suggestion. You can articulate what you want without precisely knowing how you want it. Get the results from the model and then iterate and provide feedback. And that idea of kind of working out loud, obviously, you know, via a text interface with a machine, but you're still kind of working in a slightly different way than you would maybe just on your own in front of a laptop or a terminal or a phone. So that kind of co-creation, that idea of a copilot for most people in most jobs is a pretty different way of working than, than they would. Of course, you're used to working with people as part of a team, but that's not quite the same thing when you're producing an output individually. And so that idea of how you engage with an agent in that way is pretty new. And I think that's something that I think will shift very, very rapidly, just as we saw children and teenagers almost overnight become completely familiar with how to use an iPad and an iPhone. It was just a very, very rapid shift towards something that was clearly very useful. I think that as we see these tools and these models get baked into education training, which they absolutely will do quite quickly, I think we will very quickly see people starting to work in a different way by using these tools as a co-creator, co-collaborator, co-pilot. And that applies to whether you're a software developer or a data scientist, a copywriter, a marketer. You know, these are all places where people will need to have that particular skill. More broadly than that, yeah, you know, who knows? Because these models have so many uses, so many use cases, the way that we're going to use them in different places is going to look different. And so the skills that people will need in terms of how they interact with them, or maybe don't, you know, is, is, is kind of up for grabs, I think.

What would be a good example of a problem today that is really like facing us now that you would be advising a client on? For example.

Increasingly, if we want to use these models to play an important role in, let's say, a risk management process or a customer service process, then there is a real risk that if the model gets an answer wrong, but that looks plausibly correct, that could present a genuine risk, potentially a safety risk to, to an operator, or a commercial risk, right, by giving incorrect information to a customer. And so it doesn't need to be existential to be problematic and dangerous from the perspective of a business or a consumer. And so when we think about like hallucination, that gets you into the kind of questions about the types of use cases that you want to use these models for. So even just being selective about where you use them is the first question really when it comes to kind of thinking through safety. And then clearly how you put in place the right processes, safeguards, validation, human in the loop around the large language model such that you are not just allowing the, the model itself, however convincing and plausible the outputs look, to kind of autonomously run a process end to end. But you, you're actually making— using the powers of the model, but also using the kind of validatory process that a human can play to ensure that the outputs are correct, high quality, easy to understand. And are not presenting an undue risk. And so that a lot of what we're doing in that safety space with large language models is actually thinking through those kinds of problems as opposed to thinking about whether or not there's a risk that they might escape into a kind of an uncontrolled runaway AGI type environment.

I'm Sam Gregory. I'm an executive director at a human rights network called Witness, and we're going to be talking about deepfakes and AI today. So I probably need to sort of back up and explain why a human rights network would, would care about this issue and WITNESS was created actually almost 30 years ago after an incident of police violence in the US, the Rodney King incident, which was when a Black American man was assaulted by the police and it was captured on camera. And basically over the last 30 years, what we've done is evolved with the evolution of how both professional human rights defenders, but ordinary people capture video, tell narratives about critical human rights issues and civic journalism issues all around the world. And what that's inevitably meant over time is you also have to really defend the integrity of those accounts, right? They get challenged. People say that what was filmed isn't true. They claim that it's been manipulated. They claim that even the underlying truths didn't happen. And so right now, WITNESS is an organization that works across the world. I have colleagues on 5 continents. And a big part of our work is literally working very closely with human rights activists who are filming in protests, documenting war crimes in Ukraine, filming land rights abuses in Brazil. But what we learned over time was that you also had to really think about kind of the tech infrastructure that set the terms for how, you know, an ordinary person could create images that could be trusted, could share videos that people would believe. And so about 5 years ago, we started to see the advent of what at the time was known as deepfakes, these ways in which, you know, you could swap, you know, someone's face or manipulate a video in increasingly sophisticated ways. And so we launched an initiative called Prepare, Don't Panic, which was I think the first one in the world to really say, how do we start to prepare for this potential threat, but do it in a really inclusive global way that centered people who actually already have experienced similar harms, right? So if you're a human rights activist, you know, in a country like Kenya, probably already someone has tried to fake your image in a photo. Someone has already said that your story is a lie. Someone has already tried to disprove your account. So let's start with those folks to think about how we start to structure an approach that might also be super relevant for everyone else in the world who is going to face the harms of these as they get more widespread.

I think it's really interesting that people think that they can spot deepfakes to an extent these days because in a funny sort of way, we've been educated but to an extent. So everybody thinks of the archetypal sort of Tom Cruise deepfake and, "Oh, that's a deepfake, isn't it?" But as you point out, it's much more widespread and at much greater scale. I wonder if you could take us through some of the types of— I don't know what to call it really— manipulation, I suppose, of all kinds, and where you see it cropping up most.

Yeah, so, yeah, deepfakes is a confusing word because I think for most people, yeah, it kind of summons up the idea of the face swap, right? You know, you swap Tom Cruise's face onto the face of a very talented mimic, and you have those incredible ones we've seen online. But that's really only the tip of the iceberg. And in some senses, that's also the part that is still hardest to do, right? It's still really hard to really make a completely convincing deepfake of, say, a well-known person in a way that really fools people. So, you know, it's actually the Tom Cruise example is quite a good one because under that you have a really talented professional impersonator of Tom Cruise, Miles Fisher. You have a really great company that knows how to do visual effects around it. And, you know, it's pretty convincing. But that's hard for perhaps the average person to do. And I think it's worthwhile stepping back and saying that like deepfakes are part of a field that people talk about sometimes as synthetic media. And it's part of this bigger field of AI or intersects to this field of AI called generative AI, which is sort of the buzz phrase people have heard like in relation to like ChatGPT or tools like DALL-E and Midjourney. And this whole universe allows a bunch of things. So, you know, at the very simple end, it's about things that have got much easier in the last year or two. So, you know, you can take a photo and you can animate the photo, right? And many people have done that, right? Like I did it with my great-grandmother. I colorized an old photo and then I made her, you know, move and say a few words, right? And that's got very easy because it no longer requires you to sort of train it for a specific person. It just often requires one image to do it, right? So these are some trends that have happened as well as to make it easier and not require you to have lots of images of say the person you're shifting. So the very simple end is, ways you can take an image, you can make it move. The kind of commercial things we've seen in a tool like Adobe Photoshop, you can do things that you used to be able to do with photo and you can now do with video, like remove a moving object in a video in the same way you could pull out something you didn't like in a photo using Photoshop before. So those are really simple things that have been commercialized. And then you have things that are really on the cusp of getting much more mainstream. And often people talk about things like lip sync dubbing and puppetry, which are these ideas of you can either make someone look like they said something they didn't did. So you match the lips of someone to either, you know, a voice mimic or maybe a voice clone, and you make them look like they said something they never actually said in reality, right? And that's something that is being commercialized very broadly. And I can talk about the risks of that in a second, because obviously that's actually more subtle in some ways than a full face swap. It's, you know, like you just see someone saying something and you're like, oh, they said that. So that's been happening. And then there are other ways you can do this sort of puppetry, right? You can make someone move in a way they never did, right? You can make someone look like, you know, in a funny way, like they're dancing, you know, and they never did, right? So there's a whole range of ways of sort of puppetry that are getting much easier and they're often being combined. And I sort of pointed to this by talking about audio with it getting easier to fake audio or to clone audio, right? Which is also something that's progressed pretty rapidly, right? You know, you can get quite short clips of someone's voice and then create a voice that is close enough that it might fool someone. And that sounds close enough to the types of environments we operate in, right? Like a phone call or maybe a noisy background that sound more realistic than kind of that sort of clean cut, like it's a computer voice against, you know, this very neutral background. So right in the sort of center of sort of deepfakery and synthetic media, you've got things like lip sync dubbing and these fake audio. That are in some ways perhaps some of the most threatening areas right now. And alongside that, you've got another big set of sort of commercial industry, but also much more available tools that are being used in business to do things like create an avatar of a realistic looking person, right? Like maybe looks like a news anchor saying something. It's not a real person, but it looks like a real person. And that sort of leads us into perhaps the final area of what's been happening, which is really closely related to the this area of so-called generative AI, which is the ability to create new pieces of media from existing data of events or people that never existed, right? And the place people are familiar with that over the last few years is probably, you know, encountering fake faces, right, of people who never existed. There's a website, thispersondoesnotexist.com, where you could just, you know, create endless faces of realistic-looking people who didn't exist. What's changed has been these generative AI tools, and the best known ones are like DALL-E, Midjourney, the tools built from an open source set of code called Stable Diffusion. And what those allow you to do is something known as text to image. And text to image is basically you write a sentence and it produces an image of that sentence and then potentially endless variants of that. So I might create a sentence like, as we saw last month, something like, I'd like a photo of the Pope in a puffer jacket, make it Balenciaga, right? Right. You know, and then you can generate this, and this is a big shift, and there's lots of implications of the shift with text-to-image and soon text-to-video, because it makes it much more accessible, much easier to do, to create realistic-looking images of often real people doing things they never did.

Yeah, it's fascinating, and there's so much of it around. It's interesting you mentioned the Pope, actually, because that image There's something very specific about that image. It's almost like it's that thing of something being telekinesis possible. We know it's not real, but we kind of hope it is real and we'll treat it as real. And I think that's some of the reason this stuff gets passed around in such a widespread manner, even if we're not sure, or maybe we know, but it's a good laugh, isn't it? Or it's sort of, it's, it's a joke to pass around, but it's not really a joke if you're on the other end of it. And I wondered what your point of view was on who's responsible, because I remember— was it— it was pre-pandemic, so it must have been like 2019, something like that, when Macron was tweeting that photo image of the Brazilian rainforest and attributing it to the wildfires. And actually afterwards, I think it was about 3 days after, I think people were saying, 'Hang on a minute, this photo image is 16 years old and it's nothing to do with what's going on now.' Now, I'd love your point of view on whether celebrities, politicians, people like that are just as responsible or irresponsible to pass this stuff around as like the everyday person because they've got the big accounts and they're often, they're more trusted.

Yeah. And I think I want to pick up one point you just said, Tracey, which is actually about not sort of dehistoricizing what's happening here, right? Like often when we talk about new tech, it's like, it's completely new. We don't think about who might have had experience with this in the past, who might think about the solutions, the ways we've dealt with it. And The analogy you're drawing is actually like in the work we've been doing at Witness probably for 15 years. And I often date it back, for example, to the Syrian conflict where we saw this happening a lot around human rights imagery. People shared these so-called shallowfakes, which is like someone just takes an image, they miscontextualize it, and they don't always do it maliciously. Sometimes people are just trying to represent, you know, an issue that they care about. They don't have exactly the image and they share it, you know, in order to sort of demonstrate their affiliation to something, to try and incite attention, to try and promote a cause they care about. And I've had conversations with folks who've shared images like this in quite extreme human rights crises where they talk about, I knew that it wasn't exactly what it is, but I knew that it represents that, right? And so, you know, and then you also have people who maliciously share because they're trying to incite hatred or do, you know, or deceive people. But it's a really complicated mix of motivations around how people share imagery and what it represents to them. And I think we have to understand that when we look at how people share shallowfakes and, and emotional real images, and of course deepfakes, right? And so sometimes, as you say, I think people may know they're sharing something that's manipulated. People often talk about the liar's dividend, this idea that you kind of share a footage, you claim that something real is false, but actually often it's just people leaning into what they want to believe. And I think that's coming to your question around the role of influencers and celebrities, right? They have a tremendous reach, right? And they sort of set the tone about how people share. And so I think there is a responsibility that sits with them I think the challenge is that probably for many celebrities, they're perhaps in some ways as ill-equipped as ordinary people to know increasingly how to discern it. Like, I remember Chrissy Teigen, right, referring to the Pope video, which is like, oh my God, I got fooled, right? Or words to that effect in her Twitter account. And, you know, and like, it's— it seems in some ways unfair to expect that she would be able to discern, for example, you know, is she looking forensically at each image to try and work out if the Pope has his ring on the wrong hand or has 6 fingers on one hand? You know, like, so, so I think there's both. Yes, we place responsibility on people who have greater reach to be responsible about how they share, but we also need to be very careful we don't place responsibility on this problem on essentially the end consumers, distributors, users, because we're least well placed to be able to do this.

So let's talk about some of the solutions then. Maybe tell us some of the use cases you've had, or maybe it's working with activist groups or journalists. Take us through the landscape with some of the solutions.

So the solutions, I think we really have to think across a spectrum, right? It's— there's not one silver bullet here. It's a classic phrase everyone says, but it's absolutely true here. One thing to think about is it needs a responsibility that goes across kind of a pipeline of how we make media, right? So you can't just blame the end consumer, right? So I think a lot of what's happened around, you know, the fake Trump photos or the fake Pope photos is people saying, well, surely you could have seen that the Pope always wears his ring on his left hand or his right hand. I can't remember which way it is, like the papal signet ring. Or surely you know that hands don't do well, or surely you know that you can't write text well in these types of tools. And the problem with that is it both assumes that people should have this kind of forensic mindset of looking at any image they see, which is unfair, right? I don't look at every image in my timeline or on TikTok with that view. And the second is that those clues go away over time, right? We know that those clues are just the sort of the failures of the current algorithm. So even like, for example, this tip that hands don't look great in, images created with these text-to-image isn't really true with some tools. And it's getting easier to write text in images and deepfakes didn't used to blink, but now they blink, right? So when we tell people these things, we are giving them false clues that won't last, right? And that's actually very harmful 'cause people remember stuff. People love really easy like heuristics, right? Like, oh, deepfakes don't blink. When in fact, you know, the moment that was shared as a tip, people started making deepfakes that blinked. You know, the hands, people want to improve hands, you know, as creators of these tools because it's an obvious thing to try and improve. It's a technical challenge to do that. And, but those ideas stick in people's heads. So we can't start with blaming the users. We can say you need to have, we need to double down on media literacy and say, look, you know, the same tools of like stopping, thinking if there are other places that would corroborate a story you're reading, seeing if there's an original image that this has been manipulated from, right? Those are classic media literacy strategies that we should still have. But the point is those need to be supported by the pipeline that comes before that, right? So, you know, if you— and what I mean by that pipeline is there are people who build the underlying models for this, literally the kind of ways in which the technology, the sort of foundations the technology is structured on. There are people who build tools on top of that, and there are people who create distribution, right? So if, for example, you want you or I to be able to understand that a piece of media was synthesized, you might want to have a watermark that's placed at the model stage that is visible all the way through. You might wanna have what's known as authenticity and provenance infrastructure. That's ways in which you can show that a piece of media was edited in a particular way, was distributed in a particular way that you could look at and know, oh, I know that this was, you know, a photo that was shot, you know, on the streets of Chicago last year. And then someone had did a face swap in it and then they shared it on TikTok, right? You know, so those are the types of signals that could help people make decisions on this through a pipeline that goes across this. And there's some core principles you'd want in those types of technical solutions that are around, yes, disclosure, but also consent and privacy that you need to build into how you do that, right? So those are the kinds of sort of signals we might want across our kind of media pipeline so that it's not just us guessing when we see a piece of media. And I think it's interesting. I think we can also think about that as being part of sort of the next wave of creativity in creating media. It's like I often, look at the types of videos I have in my, you know, TikTok For You page or in my following. And if you look at that, you already see that media is much more visibly constructed than say it was 10 years ago, right? You see like someone's got an AI filter, they've stitched it with something, they've, you know, they've edited in particular ways. So actually, I think we're increasingly comfortable with knowing that our media is very constructed. And so actually we can be much better at saying, actually, let's just be much more transparent about how media is made. And that might be, this was made with AI, this was made with a mix of AI and real footage, which is gonna happen all the time. This was made based on this dataset, right? Like I think we can live with a bit more complexity, but to do that you have to have the pipeline. So that's one set of solutions as kind of like the technical ways you can do this. You know, the other is like, what are things we do not allow and we do not permit to happen in these types of systems because we believe they're illegal or in a, within democratic systems otherwise, right? So, you know, should it be legal to create non-consensual sexual images of women without their consent and distribute them, right? Like, that's a place where it's like, we should have legal safeguards that, you know, make it very hard for people to build tools that are oriented towards that, that criminalize people who do that in a commercialized way or do it in a way that is used maliciously. And again, there's a whole debate around how to think about intent or not in these ones,, right? But like, there are places where we can say, look, we need to really get strong legislation that addresses something that we agree is a harm. You know, similarly, there are some laws that are starting to happen around election usage, right? Under what context should you be able to use this in, you know, in elections and those types of contexts? So there's a couple of places where this is happening. You know, I think that's all against a backdrop of like the bigger picture of AI legislation, right? Like, what are the big safeguards on transparency on this? Like, do we know what's happening under the hood of these models? Do we know that there are gonna be responses if there are widespread harms? Because in some sense, what I'm talking about here are individual media items, right? How do you recognize the Pope in the puffer jacket? But there's a bigger issue here, which applies to both images and video and text, right? Which is kind of this idea that maybe our whole information ecosystem is gonna be polluted with fake images, fake videos, some of which are hallucinations. This idea that like, Sometimes these systems create false facts or false information that looks like facts, right? And so that's a bigger issue of like the bigger sort of ecosystem that can't be addressed at every individual content item. That's a part of it, but it's a bigger picture of like, well, actually how do we wanna regulate these whole systems and what are the safeguards when a whole system starts creating information that's deceptive, false, you know, and undiscernible, right? So I think we have to think at multiple levels like the consumer, the pipeline of the technologies, the individual content item, but also this bigger picture of kind of the ecosystem around us.

You see, as I hear you talk, I'm thinking how much, what percentage of media experiences will be real in the future? I mean, will any? Because it could be that the vast majority are either exactly what you've just been explaining, these imagined futures and speculative fiction, in a sense. And also, they could be the suppression of information, which again creates a misrepresentation of reality. I mean, I think that's one of the things I'm most worried about actually, the suppression and the missing information, not misinformation. We've seen it through the elections in recent years. We've seen, you know, people disappearing out of photos. Anything that can be digital can be disappeared. And I really do think it's is a big problem because it's not like you can access it. It's not there, so you can't access it to even have a debate about it. I wondered what your thoughts are on that.

Yeah, I think there's a number of trends, including the one you're describing, that are important to think about when we're thinking about how much is real, how much is fake. And I also want to push back a little bit on the real and fake because I think we're going to be in a world that is much more sort of mix, and we'll have to just be better at interpreting that.

Well, it's all one thing eventually. It's all a real-fake, genuine-fake world. And we're creating those ourselves with profiles. They're us but not us in a way. And we're trying to navigate this liminal space. So I know what you mean. I do agree. It's just the only way I can describe it at the moment is this distinction.

I think it's— what I'm sort of thinking about there is like, you know, when we create a photo, like we already use AI to like, you know, improve our faces, make ourselves smile better, put the bokeh effect in the background. So like sort of naming that, because I think often we get these statistics like 90% of things are going to be synthetic by 2026 is a common one that's named. I'm like, I'm not sure how helpful that is 'cause it sort of depends. Is it the 90% we care about or is it the 90% that we want to be synthetic, right? Like if it's 90% of my TikTok feed is synthetic, I'm kind of already expecting that and I kind of want that and I'm enjoying it. So I think like just knowing that numbers and the sort of the real false is gonna be complicated and also will be used against us, right? 'Cause it'll be like, well, you can't believe 90% of stuff when in fact we might want 90% of stuff to be not exactly real in some sense. Yeah, so, but I think that there's some trends that are important and you're naming one is like, will it be easier to kind of erase things? And I think it's true that like this allows us to erase pieces of content, right? It's easier to manipulate discrete content items. I think there are some other things that are sort of around it that go to the same worry you have, right? One is the idea that this also makes it easier to personalize realities for particular people, that very much this is about like personalized information. You know, you can already see it with like the ad, sort of the ideas for using generative AI in ads are about personalized ads, right? And that has a real pernicious effect 'cause you can personalize in ways that reinforce your biases, that put you in even more of a bubble of one, right? And so if we're trying to think about how we share information, how we engage each other on a shared platform of actual information, the more we're personalized, there's a risk that we're sort of, we're reducing that. The second is I think that one of the sort of ways in which this potentially changes our information environment, it's particularly to do with some of these trends that are happening with the accessibility of these tools is the ability to create volume, right? And so, you know, we could think about that in, you know, in the ways we might think about very organized strategies around trying to deceive people, right? There's this idea of floods of falsehood that comes up in, you know, disinformation discussions, which is like you can overwhelm an information environment with lots of contradictory accounts and people just throw their hands up in the air because they have no idea what's true. So volume is easier in this environment. I also just think there's a volume creation thing that makes it harder possibly to find the things that you might want to be finding or you should be seeing. And we definitely hear that from human rights activists. They're like, I can't create 60 different accounts of that incident that happened. There is only one account, and it's the important one. But an opposing actor, be they a corporate actor or a government actor, could create 100 different contradictory accounts. They have no ethics about this. They can create 100 100 stories and I have just my one story that really matters, right? And so, and they might also try to erase that story, right? They might try and release a version of that story that has, you know, removed, you know, the police officer who is assaulting a protester, right? So all of these kind of come together. And meanwhile, lots of other people might be in these environments which are very personalized and bubbled where they really are just living in a mixed reality world. And other people have thrown up their hands 'cause they're like, you know, so many false and true accounts I don't know what to believe. And, you know, I'm listening to, you know, a ChatGPT or equivalent, and I know it hallucinates. And so you can't even tell when it tells you an answer whether it's true or false, right? So I think there's a lot of factors here, and this is really about the bigger information ecosystem where we need to go, well, how do we think about those implications, you know, both to protect individual and maybe, you know, group actors who really want to deceive us or have an interest in maliciously deceiving us or, avoiding accountability for their actions, but also just more broadly, like, how we may all be in a place where we are more confused, less confident in what we see, and then these critical voices who we really perhaps should trust are the ones who have the hardest time breaking through the chatter.

It's good to spend as much time as possible, I think, in the future with real human beings, just to, like, keep re-anchoring ourselves into, like, the physical— I'm not gonna say real the physical world, I think, is gonna be more and more important. Just as we wrap up, and you've covered so much, and it's all so fascinating, I wonder if you want to just tell us a little bit about what's on your horizon. I could ask you, where will we be by 2030? But I think it's impossible, really, to say right now, and things are moving so fast, and as you say, the speed and the the scale of everything. But what are some of the trends you think are reaching out into the next few years that we should— I mean, I'm really worried about voice. I mean, I'm really worried about— I've seen what ElevenLabs and the like are doing. I think I spoke to the bank the other day. I rang up the bank, and I go through, obviously, voice biometrics. They're like, will you say your name, say your address? And I said to her, this— I really do not feel confident about this technology anymore. And I started to talk to her about this. And she said, 'Oh, it's okay. Our systems, you know, you can't override our systems.' Like, the naivety of it is just incredible. So I'm quite worried about voice, but I'd love to know what you're worried about and looking at on the horizon.

Yeah, so the first thing to say is it's quite hard to predict, like, long trajectories, even medium to sort of even medium term, right? I think the sort of generative AI explosion of the last 6 or 8 months probably came faster than myself and others expected. I'm in the US context and people are always asking, what are we gonna do in the 2024 elections? I'm like, that's 14 months away or 16 months away. I don't know exactly where we'll be. I think there are things we can point to and I agree. I think there are technology trends where you can see trend lines. So audio has got a lot easier to do and we're seeing it used in very common everyday context like you're describing. Someone fooled actually the Australian National Biometric System recently We're seeing scam calls being used. We're seeing swatting calls in the US where it's an automated voice, right, calling to do swatting on folks. So audio is improving rapidly. And I find very worrying also looking globally, we've seen a lot of misuse of audio. It also lacks all the sort of semantic clues around it, right? A video, like there's so many things you could look at to see if it's wrong. While audio, you're just kind of listening and you're going really hard, is it a little too electronic? Right? There's just no surrounding clues, particularly for example, also like when it's, for example, shared in a WhatsApp group, as in many parts of the world where voice messages in WhatsApp are like, you know, how people communicate. So audio's improving rapidly. These text-to-image and text-to-video tools have been improving extremely rapidly, right? Like I was even describing like the progress in the image generation that's gone, like in the way to generate hands. Text-to-video still looks kind of funny if you've seen that, right? Like, you know, Will Smith eating spaghetti, but it's moving fast. I don't think it's like at the speed where we're gonna have like full-length, you know, complex deceptive videos. All that's super soon, right? Cautious about my words, but those have advanced really rapidly. You know, the ability to create realistic-looking avatars that say things advanced really rapidly. So all these tech trends are advancing. It still remains somewhat hard to do a super convincing, face swap deepfake in a kind of real-time context. So if something happened this afternoon in New York City that was like where I wanted to face swap someone into it and have it ready by the next morning, I would still be a little cautious about whether that's possible. So that's the place where it's still like, you know, where a lot of people's fears sit. But actually probably the realities is the day-to-day audio, text-to-image, video that never existed, and the non-consensual sexual deepfakes that continue to proliferate, get easier and easier to use. So that's sort of the tech trend. The trend lines on things like regulation, suddenly everyone's paying attention to this. You've got proposals in the UK, you've got proposals in the US, you've got the EU pretty close to final stages on the EU AI Act. China just passed or, you know, sent for comment, but it'll pass a set of regulations around this, you know, and I think there's, you know, we'll see regulation in the next couple of years. And the critical question is, will it address address adequately this pipeline responsibility, you know, and will it create collateral harms? Like, I spend a lot of my time working with my colleagues who are around the rest of the world, and often you see legislation that's kind of copycat legislation, you know, to suppress journalistic freedoms, to suppress satire, to suppress free speech that is used in response to these threats. And we've even seen it, we saw it around COVID, we saw it around, you know, the fear of deepfakes, right? So I I think regulation is important, but it really needs to go to the core of the problem, right? Like the transparency, the regulation of these models, the real attention to very specific harms, the accountability of tech developers. And it needs to really think as it sort of proposes infrastructure solutions like this authenticity solutions that it doesn't do that while compromising privacy, for example, right? So I think the tech trend is, and the sort of the regulatory trend is action, action, action. And the question is, how do we make sure that's actually addressing the problems, it isn't creating the harms? The bigger picture, like I worry very much around this kind of just both the individual content undermining critical voices in our societies and attacking ordinary people and also critical truth tellers, but also this broader undermining of kind of our epistemic trust, right? Like this belief that we have, and I think you can approach that in multiple ways, let's say from our own, you know, witness perspective, we're saying you have to fortify the truth. And some of that might be tactical, literally like Film360 in a critical event. Some of that might be infrastructure, like having a tool that can show how a piece of media was made. And some of that has to be, you know, policy side, right? Like protecting people's right to film in public spaces is important, right? So you actually have the literal documentation. And it's also gonna be like the types of AI regulation that actually structure a space that still, that doesn't suppress speech, that allows people to be trusted, that doesn't undermine this shared ecosystem that we still need to aspire to. I don't think we can give up on the idea that we should be able to communicate with each other in a shared space.

I was just thinking about when you talk about media literacy, and I would say it's even more than that, in addition it's media forensics. Like in terms of like almost putting it on the curriculum, like it's almost one of those 21st century skills that people are going to definitely. But the question I have in my mind is, are there enough people to teach it? Where do we start? This is—

it's actually, I realized I didn't talk about this because one of the things we've spent a lot of time working on in, in Witness has been around detection. So like, I've been talking a lot around kind of like media literacy and authenticity, but like, obviously people also ask like, how do we detect this? Can we run a tool that shows you something's manipulated? And there's sort of multiple problems around that. One is that the tools aren't they don't work very well on kind of social media video or photos, and they're not widely available, and the expertise around them is not widely available. So, you know, we often talk about an equity gap around media forensics, which is that the people who need it most don't have access to it. There's just, you know, it's very much concentrated in academia, law enforcement, you know, intelligence agencies, not in civil society, not in most journalism. And so I think we need to think about, and this is a big structural challenge, how do we have more people who have expertise in media forensics, including in spaces like civil society, like journalism, you know, not just in places that may have mixed motivations for when they wanna disclose whether something's fake or not and, you know, and are controlled by the state, right? Versus, you know, civil society and the media. So we wanna do that. I get very conflicted about media forensics. It's interesting 'cause I think part of what we're doing, if we try and encourage a sort of, people talk about the forensic turn, in some of the work that people do, which is like looking really closely at images and video and trying to scrutinize them. And you see it around Ukraine, people like going like, I'm just going to look at every pixel and try and work out whether that's a reflection in that mirror and was that real, was that fake? And the danger in that is like, it probably isn't that productive for most people because they don't have the skills or the tools, but it encourages a skepticism that is— that can't get us to a place. So it's a really interesting balance. We want to have more availability of media forensics. We want tools tools that are interpretable, that are available to more people. But we probably don't want to switch media literacy for media forensics for individual people, because actually probably in most cases it can be better to say, actually, is this likely to be credible? Do other sources confirm this? You know, was this manipulated from an original versus trying to like look really closely at the pixels? So it's a complicated landscape, like as we think about media forensics, there's definitely the gap you point to, the sort of bottleneck and we're going to need more of it. At the same time, we probably want to sort of get a right balancing act about how much we encourage individual people to be media forensicists— doesn't sound right, does it? Forensic experts versus, you know, so it's an interesting balance.

Thank you for listening to The Future Review hosted by me, Tracey Follows. Check out Check out the show notes for more info about the topics covered in this episode. Do like and subscribe wherever you listen to podcasts. And if you know someone you think will enjoy this episode, please do share it with them. Visit thefutureofyou.co.uk for more on the future of identity in a digital world and futuremade.consulting for the future of everything else. The Future of You podcast is produced by Big Tent Media.